Graphics Reference
In-Depth Information
Output Block Size
a
Raw Input Format Bits per Pixel
HDR RGB+A
Bits per Pixel
64
4
×
4
8.000
5
×
5
HDR RGBA
64
5.120
HDR RGB
48
6
×
6
3.556
HDR XY+Z
48
8
×
8
2.000
HDR X+Y
32
10
×
10
1.280
RGB+A
32
12
×
12
0.889
RGBA
32
3
×
3
×
3
4.741
XY+Z
24
4
×
4
×
4
2.000
RGB
24
5
×
5
×
5
1.024
HDR L
16
6 0.593
4
×
6 5.333
8
×
10 1.600
12
×
10 1.067
a
This is by no means an exhaustive list of available block sizes,
merely the square/cube block sizes to show data rates, with a
few nonsquare examples.
6
×
6
×
X+Y
16
LA
16
L
8
Tab l e 6. 1.
Bitrates of raw image formats.
Tab l e 6. 2.
Bitrates of ASTC output.
Regardless of the blocks' dimensions, they are always stored in 128 bits, hence
the sliding scale of bit rates.
6.3 Algorithm
Each pixel in these blocks is defined as a quantized point on a linear gradient
between a pair of boundary colors. This allows for fairly smooth areas of shading.
For blocks containing boundaries between areas of completely different colors, the
block can use one of 2048 color partitioning patterns, which split the block into
different designs of 1-4 color gradients. (See Figure 6.1.)
These blocks are algorithmically generated, and selecting the right one is
where the majority of the compression time goes. This technique allows a block
to contain areas of completely different hues with arbitrary shading or multiple
intersecting hard-edged patches of different tones. Each block defines up to four
pairs of colors and a distribution pattern ID, so that each pixel knows which
of those pairs it uses to define its own color. The individual pixels then have
aquantizedvaluefrom0to1tostatewhere they are on the gradient between
the given pair of colors.
1
Due to the variable number of bounding colors and
individual pixels in each 128 bit block, the precision of each pixel within the
block is quantized to fit in the available remaining data size.
1
Indices are encoded at variable resolution using a scheme developed by researchers at Ad-
vanced Micro Devices, Inc.
